Understanding Google A2A in one article: What is it? Why can you team up with AI Agents?

The curl is numb...

Last night, Google open-sourced the first standard agent interaction protocol, Agent2Agent Protocol (A2A for short), at the Google Cloud Next 25 conference .

Actually I saw this message before going to bed, but I really wanted to sleep..

When this thing came out, it felt like the entire AI Agent community was shaken.

Why? Because it may really change the way intelligent entities interact overnight and completely break down the “system silos”!

I feel like the concept of AI Agent is everywhere these days.

Various cool Agent applications are also emerging.

But do you have a feeling that these agents seem to live in their own little world?

When you ask the Agent of application A to do something, it may not know how to greet or exchange information with the Agent of application B.

It's like you hired several experts (Agents), but they don't know each other and can't understand each other's "jargon". You want them to work together to complete a big project? Difficult! Low efficiency!

This is the pain point we face today : collaboration barriers between agents and information islands between systems .

But! Just last night, Google took action and wanted to dominate the market - it open-sourced the first standard intelligent agent interaction protocol: Agent2Agent Protocol, abbreviated as A2A !

What is this A2A? What does it want to do?

Don’t worry, today Jiamu will use plain language to help you understand it thoroughly!

A2A: Mandarin in the AI ​​Agent World

We can substitute that the current AI Agent world is like an ancient "World Expo".

Each country (representing a different company or platform) sent its own agent.

Each of these envoys has unique skills and can handle various affairs, but the problem is that they speak the dialect of their respective countries and follow the etiquette norms of their respective countries.

You want the Qin envoy (such as Salesforce Agent) and the Zhao envoy (such as Workday Agent) to work together to do something, such as verifying an employee's sales performance and corresponding HR information.

They may have to find a translator first and learn the other party's communication style, which is extremely inefficient and may even lead to mistakes due to misunderstandings.

This is the problem of siloed data systems and applications  .

The A2A protocol is like a set of "universal language" (standard communication language)  and  "universal handshake" (standard interaction method ) jointly formulated by Google and a group of big players (there are more than 50 well-known companies in the first batch! There will be more in the future!) .

Let's take Jiamu as an example:

• A2A ≈ "Mandarin" + "Standard Interface" in the AI ​​Agent World

With the A2A standard, no matter which "country" (platform/supplier) your "Agent" is from, as long as you learn this "Mandarin" (follow the A2A protocol), you can communicate smoothly with other messengers who can also speak "Mandarin", exchange information securely, and coordinate actions.

Just like having a USB port, different brands of USB flash drives, mice, and keyboards can be plugged into a computer and used. The goal of A2A is to enable different "brands" of Agents to work together in a plug-and-play manner  ! (I will talk about the relationship with MCP later, wait~)

So, the core definition of A2A is:

A2A is an open protocol that provides a standard way for agents to interact, enabling them to collaborate with each other regardless of their underlying framework or vendor.

Why is A2A so important?

You may ask, is it really that exaggerated to just make an agreement?

Jiamu tells you, this matter may not be simple!

Often when a new thing emerges there will be a brief period of vacancy in its ecological niche. At this time, if someone takes the lead in initiating an agreement and setting standards, that means they have secured this ecological niche.

Think about when Google took the lead in developing  Android . Before Android came out, there were many different mobile phone systems, including Nokia's Symbian, Microsoft's Windows Mobile, BlackBerry's BlackBerry OS...

Application development has to adapt to various systems, and the user experience is also fragmented. Google joined more than 80 companies to form the "Open Handset Alliance" and launched the open source Android system. The result? It directly unified the mobile operating system world (except Apple iOS).

This time, Google open-sourced A2A, and immediately attracted   a number of industry giants such as  Atlassian, Box, Cohere, Intuit, Langchain, MongoDB, PayPal, Salesforce, SAP, ServiceNow, UKG, Workday, etc., as well as top consulting companies such as Accenture, BCG, Deloitte, KPMG, McKinsey, and PwC  to support it...

Doesn’t this campaign have a bit of the flavor of Android back then? ?

Interpretation:

Google's goal in this move may be very ambitious, and it wants to occupy the position of Big Brother:

1. Unify the Agent ecosystem:  Solve the current problems of fragmentation and confusion in the Agent market and establish a universal "rule of the game".
2. Break down data silos:  Enable agents deployed on different systems (CRM, ERP, HR, project management, etc.) within the enterprise to collaborate seamlessly and truly unleash the power of “combination”.
3. Accelerate the implementation of Agent applications:  With standard protocols, developers can more easily build cross-platform Agent applications, and enterprises can more conveniently integrate and manage Agents from different vendors.
4. Seize the entrance to the future: 2025 is undoubtedly the big year for Agent. This is not only a technical matter, but also a layout for Google's business ecosystem and future strategy!

Moreover, at this conference, Agent was clearly the focus of Google. In addition to A2A, Google also launched:

• Agent Development Kit ADK (Agent Development Kit):  Following the example of OpenAI, it lowers the threshold for Agent development.
• Internal testing tool Agent Engine:  used to test and evaluate Agent.
• New Agent Marketplace:  Create an “app store” for Agents.

With this combination of punches, Google's determination to make something big in the Agent field is obvious!

Five design principles of A2A

In order to make A2A, the "world's universal language", easy to use and reliable, Google and its partners followed five key principles in its design.

These five principles determine A2A's "character" and capabilities:

Principle 1: Embrace agentic capabilities

• Original text:  A2A focuses on enabling agents to collaborate in their natural, unstructured modes, even if they do not share memory, tools, and context. Google is enabling true multi-agent scenarios rather than limiting agents to being a tool.
• Jiamu Plain Language:  A2A does not want to turn Agents into "tools" that can only execute fixed instructions. It hopes that Agents can collaborate like a real team, and can communicate more freely and closer to natural language. Even if their internal "memories" (data) and "toolboxes" (capabilities) are different, they can understand each other's intentions and complete tasks together.
• Example scenario:  Imagine two agents are working together to write a report. Agent A is good at collecting data, and Agent B is good at writing. Through A2A, Agent A can directly tell Agent B: "I found the latest data on market trends, the main points are as follows... How do you think it can be integrated into the report?" instead of rigidly calling a "write report" tool function of Agent B. This is more like a conversation between team members rather than calling an API .

Principle 2: Build on existing standards

• Original text:  The protocol is built on existing, popular standards, including HTTP, Server-Side Events (SSE), JSON-RPC, etc., which means it is easier to integrate with the existing IT stack that enterprises already use every day.
• Jiamu Plain Language:  Google did not reinvent the wheel. A2A uses "building materials" that everyone is familiar with, such as the HTTP protocol commonly used on the Internet (you use it when you go online), SSE (used for real-time server push messages), and JSON-RPC (a lightweight remote call protocol). The advantage of this is that it will be much easier for enterprises to connect A2A to existing IT systems without making a big fuss.
• Scenario example:  An e-commerce company may have already used HTTP and JSON-RPC for communication in its order system, inventory system, and logistics system. Now you want to let the agents in these systems interact through A2A, for example, the order agent wants to ask the logistics agent where a certain order is. Because A2A is also based on these standards, the technical connection will be much smoother, just like opening a new "smart lane" on an existing highway instead of building a new road. Low integration cost and quick to get started!

Principle 3: Secure by default

• Original text:  A2A is designed to support enterprise-level authentication and authorization, and is equivalent to OpenAPI's authentication scheme when it is launched.
• Jiamu Plain Language:  Security is the lifeline of enterprise applications! A2A was designed with this in mind. It supports enterprise-level "authentication" (who are you?) and "authorization" (what are you allowed to do?). Moreover, it is compatible with the popular OpenAPI authentication method (used by many APIs). This means that communication between agents is secure, and not just anyone can access sensitive data. It is worth mentioning that Google said "equivalent to OpenAPI", which is quite humane and does not exclude the OpenAI ecosystem, making it convenient for developers to migrate and integrate.
• Scenario example:  The agent in the finance department wants to obtain the salary information of an employee from the agent in the HR department for reimbursement approval. Through A2A, the HR agent will first verify the "identity" of the finance agent (it is indeed an authorized agent), then check its "authority" (whether it has the right to view the salary information), and only after confirming that it is correct will the data be securely transmitted. Full encryption, controllable permissions, prevent data leakage!

Principle 4: Support for long-running tasks

• Original text:  Google designed A2A to be flexible and able to support a variety of scenarios, from quick tasks to in-depth research that may take hours or even days (when humans are involved). Throughout the process, A2A can provide users with real-time feedback, notifications, and status updates.
• Jiamu Plain Language:  Some tasks handled by the Agent may be responded to in seconds, while others may take a long time, such as doing a complex simulation calculation, or requiring human confirmation in the middle. A2A takes this into consideration. It can not only handle "lightning tasks", but also manage tasks that require "marathon" execution time. And during the execution of the task, it can also continuously send "progress reports" to you (or the Agent that initiated the task).
• Example scenario:  A scientific research agent receives a task: "Simulate the pharmacokinetics of new drug X in different populations". This may take several days to run. Through A2A, this scientific research agent will:
• Tell the Agent that initiated the task: "The task has been received, calculation has started, and it is expected to take 3 days."
• During the calculation process, updates are sent regularly: "30% completed, currently processing Asian population data..."
• If you encounter a problem and need human intervention: "An exception was encountered when simulating African population data. You need to confirm the setting of parameter Y."
• When it is finally completed: "Mission accomplished! Simulation report generated."
• It’s like having a reliable project assistant who always reports to you on the progress of long-term tasks so that you can have peace of mind.

Principle 5: Modality agnostic

This is also the most practical point in my opinion.. no mode limit..?️?️??

• Original text:  The world of Agent is not limited to text, so A2A supports various modalities including audio, image, and video streams.
• Jiamu Plain Language:  Communication between agents cannot be limited to "type chat". Future agents need to be able to see, hear, and speak. A2A is designed to support the transmission of various types of data, not only text, but also pictures, sounds, and even video streams can be transmitted through A2A.
• Scenario example:  A customer service agent is handling a complaint about a damaged product. The user sends a photo of the damaged part. The customer service agent can send the photo to the technical support agent through A2A. After analyzing the photo, the technical support agent may send back a video of operation instructions to the customer service agent, who will then pass it on to the user. Through the seamless flow of multimedia information, agent collaboration is richer and more intuitive! More scenario-based

These five principles outline the powerful, flexible, secure, and easy-to-use features of the A2A protocol, and also demonstrate Google's deep understanding of future agent collaboration scenarios.

How does A2A work? Disassembling the “dialogue process” between agents

Now that we understand the design concept of A2A, let’s take a look at how it enables Agents to “chat” and “work”.

This process mainly involves the interaction between the "client agent" and the "remote agent".

• Client Agent:  Responsible for initiating tasks and communicating requirements. It can be understood as a "project manager" or "questioner".
• Remote Agent:  Responsible for receiving tasks and taking actions, providing information or performing operations. It can be understood as an "expert" or "executor".

Image description:  The left side is the Client Agent, the right side is the Remote Agent, and the middle is the A2A Protocol channel. The arrows represent the communication through A2A, including capability discovery, task management, collaborative messaging, and user experience negotiation.

A2A provides several key capabilities in this interactive process:

1. Capability Discovery: Agent’s “Self-Introduction”

• How to do it?  Agents can   advertise what they are good at through something  called an "Agent Card" . This "Agent Card" is written in JSON format (a common data format), which describes what the agent can do, how to contact it, etc.
• What does it do?  The client Agent (project manager) can quickly find which remote Agent (expert) is best suited to perform a specific task by viewing these "Agent Cards".
• Let me give you an analogy:  it's like you browse the "resume" (Agent Card) of a candidate on a recruitment website. The resume lists the candidate's skills, experience, and contact information. Once you've chosen a candidate, you know how to contact him/her and assign the task to him/her. The Agent Card is the "public resume" of the Agent.

2. Task Management: Work around “tasks”

• Core concept:  All communication between the client and the remote agent revolves around completing tasks  . The A2A protocol defines a  "task" object .
• Task lifecycle:  This "task" object has its own lifecycle (for example: pending, in progress, completed, failed, etc.).
• Task complexity:  Simple tasks may be completed immediately. For complex tasks that require long-term execution, agents can communicate with each other through A2A to keep the task completion status synchronized.
• Task output:  When a task is completed, its output is called  an "artifact" . This "artifact" may be text, data, pictures, reports, etc.
• Jiamu uses an analogy:  the whole process is like managing a project.
• The client agent initiates a "task" (project establishment).
• This "task" object records information such as the project's goals, status, and person in charge.
• During the execution, the remote agent will update the status of the "task" (project progress update).
• Finally, the remote agent delivers the “artifact” (project results).
• A2A provides a set of standard "project management processes" to regulate the work of agents.

3. Collaboration: Real-time communication between agents

• Method:  Agents can send messages to each other .
• Message content:  These messages can contain a variety of information, such as:
• Contextual information:  “Regarding the solution we discussed last time…”
• Response:  “Okay, I understand.” “There’s something wrong with this data…”
• Artifacts:  "Here is the draft report you requested."
• User instructions:  (If user intervention is required) "Please confirm this design."
• Purpose:  Through this flexible message passing, agents can work together better and complete complex tasks together.
• Jiamu gives an analogy:  It's like using instant messaging tools to communicate between team members. You can send texts, transfer files, @ someone, request feedback, etc. A2A provides Agent with a built-in "work group chat" mechanism.

4. User Experience Negotiation: Ensure the result is both good-looking and easy to use

• Background:  The "results" generated by the Agent are ultimately presented to the user. Different user devices (mobile phones, computers) or different client applications may display content in different formats.
• How?  A2A messages contain content fragments called  "parts"  (for example, a generated image is a "part"). Each "part" has a specified content type (for example image/png, text/html).
• Negotiation mechanism:  This enables the client agent and the remote agent to negotiate the content format  that best suits the current scenario . Moreover, this negotiation explicitly includes negotiation of user interface capabilities  , such as:
• Can the client display  iframes  (embedded web pages)?
• Can you play the video ?
• Can it process web forms ?
• Goal:  This way, A2A can provide the best user experience based on the user's needs and the capabilities of their device .
• Jiamu gives an analogy:  you ask Agent to generate a report.
• Agent A (client) tells Agent B (remote): "My users are using mobile phones with small screens. It is best not to give too complicated charts. Just give a concise summary and a list of key data."
• After receiving the request, Agent B generates the "artifact" according to this requirement to ensure that it can be clearly displayed on the mobile phone.
• If the user is using a computer, Agent A may say: "My user is using a large-screen computer that can display interactive charts. Please generate an embedded web chart (iframe) for me." Agent B will generate the corresponding format.
• A2A allows agents to "discuss" with each other to ensure that the final deliverable is understandable and usable by users.  ✨
  A2A working diagram

Through this mechanism, A2A provides a relatively complete and flexible framework for collaboration between agents, taking into account everything from "who to do the work" to "how to do it" and then to "how to present the results".

Real-world example: How A2A made hiring software engineers easier

Theory alone is a bit abstract, so let’s look at the specific example mentioned by Google: using A2A collaboration to recruit software engineers .

Imagine you are a hiring manager looking to hire a software engineer with a specific technology stack and in a specific location.

Without A2A:  You may need:

1. Open recruitment website A, post a job, search for resumes, download...
2. Open the internal HR system B to check if there are any suitable internal candidates...
3. Open LinkedIn C and search for people who meet the criteria...
4. Manually sort out the candidate list and send emails to suitable candidates to arrange interviews...
5. After the interview, log in to the background check service D's website and submit a background check request...
6. The entire process requires switching back and forth between multiple systems and manually copying and pasting information, which is inefficient.

With A2A collaboration:

1. Initiate a task:  You tell your "recruitment assistant agent" in a unified interface (such as Agentspace envisioned by Google): "Help me find candidates who meet this job description (attach JD), are in Shanghai, and are proficient in Python and K8s."
2. Agent collaboration (via A2A):

• Your "recruitment assistant agent" has discovered (Capability Discovery)  several "expert agents": agents that specialize in connecting to major recruitment websites, agents that connect to the company's internal HR system, and agents that connect to LinkedIn.
• It assigns tasks via A2A  (Task Management)  : "Please filter candidates based on the following criteria:..."
• Each "expert agent" (remote agent) starts working and sends the potential candidate information found (as  an artifact )   back to your "recruitment assistant agent" through A2A (Collaboration) .
• Results & next steps:
• Your "Recruitment Assistant Agent" aggregates the information received and presents you a list of candidates on an interface based on your preferences ( UX Negotiation ).
• After reviewing, you say to your agent: "These people look good. Please arrange a first round of interviews for me."
• Continue Collaboration:
• Your agent contacts the "Interview Arrangement Agent" through A2A to coordinate the time of the candidate and the interviewer and send an interview invitation.
• After the interview process is over, you instruct: "Do a background check on Zhang San who passed the interview."
• Your Agent contacts the “Background Check Agent” through A2A and submits a background check request.

Did you see it?

During the whole process, you only need to interact with your main Agent.

The complex cross-system information search, task coordination, and process advancement are all automatically completed by Agents through the A2A protocol!

Efficiency is greatly improved and experience is extremely simplified!

This is the power of A2A!

A2A vs MCP: Not a substitute, but a complement!

MCP has been very popular recently. If you want to know more about MCP, you can read Jiang Shu's

MCP: A universal connector in the AI ​​world, the next-generation standard that experts are paying attention to

Before A2A, the community (especially Anthropic) proposed  MCP (Model Context Protocol)  , which is mainly used to help LLM/Agent  connect and use tools/resources . MCP focuses more on how to enable Agent  to call external capabilities in a structured way .

So what is the relationship between A2A and MCP? Will they compete with each other?

Google’s official explanation is very clear: A2A and MCP are complementary, not competitive!

In simple terms:

• MCP: Used by Agent to call "tools"
• Scenario:  Agent needs to call a weather API, operate a database, execute a piece of code, etc. The input and output of these "tools" are usually clear and structured.
• Purpose:  Standardize the "function calls" between Agent and tools.
• A2A: For "collaborative dialogue" between agents
• Scenario:  An agent needs to discuss issues with another agent, assign tasks, pass unstructured information, and conduct multiple rounds of communication to achieve complex goals.
• Purpose:  Standardize the "application layer" communication protocol between agents to support more dynamic and human-like interactions.




The “Garage Shop” Metaphor:

Imagine an AI-driven auto repair shop with several "mechanic agents" in it.

• Where MCP comes in handy:  When the mechanic Agent needs to use tools  such as jacks, multimeters, and wrenches , it uses the MCP protocol to precisely control these tools (for example, "raise the jack by 2 meters," "turn the wrench to the right by 4 mm"). This is the interaction between the Agent and  structured tools  .
• Where A2A helps:
• When a customer (can be a real person or another agent) comes to report a repair, he or she will say to the mechanic agent: “My car makes a rattling sound.” This natural language, unstructured description needs to be transmitted and understood through A2A.
• During the diagnosis process, the mechanic agent may need to communicate back and forth with the customer : "Take a photo of the left front wheel and show it to me", "I found a fluid leak, how long has this been going on?" This process of multiple rounds of dialogue and dynamic adjustment of plans also requires A2A support.
• The auto mechanic Agent may also need to contact the parts supplier Agent : "I need a part with model number XYZ, is it in stock?" This is also a collaboration between Agents, which requires A2A.

To summarize:

• If the Agent wants to use "hammer and nails" (tools), use MCP.
• When an agent wants to hold meetings, discuss, or assign tasks with “colleagues, customers, and suppliers” (other agents or people), A2A is used.

A mature Agent application may need both MCP (connection tool) and A2A (connection with other Agents). The two complement each other!

Google even suggests that the A2A Agent itself (via their Agent Card)

Modeled as a resource of MCP, the Agent Framework can then uniformly discover and manage available “tools” (MCP) and collaborative “partners” (A2A).




The stars are shining: Who has joined A2A's "circle of friends"?

The success of a standard depends largely on how many people are willing to use it.

A2A is backed by Google, so...it has received support from a large number of heavyweight companies as soon as it was released!

Check out this star-studded list (partial):

• Enterprise application giants:  Salesforce, SAP, ServiceNow, Workday, UKG, Intuit... (covering core areas such as CRM, ERP, HR, ITSM, etc.)
• Collaboration and content management:  Atlassian (Jira, Confluence), Box...
• Databases and data platforms:  MongoDB, Neo4j, DataStax, Elastic...
• AI and Big Model Companies:  Cohere, C3 AI, Articul8...
• Development tools and frameworks:  LangChain, JetBrains, JFrog...
• Fintech:  PayPal...
• Consulting and service giants:  Accenture, BCG, Deloitte, KPMG, McKinsey, PwC, TCS, Wipro... (These companies will recommend A2A to their corporate clients!)

How luxurious is this lineup? It basically covers major players in various key fields such as enterprise software, cloud computing, AI, development tools, and consulting services.

Interpretation of Jiamu:

• Broad industry coverage:  This means A2A has the potential to connect every aspect of business operations.
• Recognition from leading companies:  These big players are willing to invest resources to support A2A, which shows that they are optimistic about this direction and are willing to participate in building a joint ecosystem.
• Strong promotion capabilities:  Especially with the participation of consulting companies, it will greatly promote the recognition and application of A2A among a large number of corporate customers.

This laid a very good foundation for the success of A2A. Google did not fight alone this time, but assembled an "Avengers" level team to jointly define the future of Agent collaboration. It can be seen that the ambition is huge. In fact, some domestic manufacturers have also made plans...

The future is here: A2A opens a new era of agent interoperability

I have discussed with Jiang Shu before, who will occupy the ecological niche of inter-agent communication?

Although we have ANP, this may still depend on resources.

Small teams may have their own limitations compared to large companies.

At this time, large companies will step forward to seize the ecological niche.

And Google's A2A seems to be opening an era of agent interoperability!

We all look forward to such a scene:

• The personal assistant Agent can seamlessly call the workflow Agent to handle company affairs.
• Shopping Agent can automatically collaborate with price comparison agent, logistics agent, and payment agent to provide you with a one-stop shopping experience.
• Agents in each business system of the enterprise (sales, marketing, customer service, production, supply chain) can operate like a highly coordinated team, share information in real time, and automatically optimize processes.

This is no longer science fiction! A2A is paving the way for this future.

Of course, A2A is still a draft specification . Although it has been open sourced and has received support from many partners, it still requires the joint efforts of the community and continuous iteration to become a truly widely adopted and mature industry standard.

Google also said that they are working with partners to  launch a version that can be used in production environments later this year .

Conclusion

The release of A2A is a milestone in the development of AI Agent.

It marks the beginning of the industry's realization that agents working alone have limited capabilities, and interconnection and collaboration are the key to unleashing the full potential of agents .

How similar is this to the early development of the Internet?

From the initial BBS and mailing lists to the standardization of Web protocols (HTTP, HTML) and the rise of the API economy, connectivity and standards have always been the core driving force behind technology adoption and innovation.

A2A may play the role of "HTTP" or "TCP/IP" in the Agent era. Of course, the promotion and implementation of the standard will take some time, and the level of intelligence needs to be improved.

But the significance of A2A is far more than a line of code or a new standard. It is more like a metaphor, a profound metaphor about the nature of connection, collaboration and intelligence .

True intelligence, especially intelligence that can solve complex real-world problems, must be born out of connection and collaboration.

What does A2A mean to ordinary people like us, especially AI practitioners and enthusiasts?

1. Pay attention to trends:  Understand the importance of agent interoperability and pay attention to the development of A2A and its ecosystem.
2. Learn new knowledge:  If you are a developer, you can start to understand the A2A specifications and related tools (such as Google's ADK, and the support of A2A in frameworks such as LangChain and CrewAI).
3. Think about applications:  Think about what new possibilities collaboration between agents can bring in your own field. How can you use this interoperability to solve practical problems and improve efficiency?

Remember, technology itself is neutral; the key lies in how we understand and apply it.

The most important thing is how we choose to walk through this door .

Should we just marvel at the miracle of technology? Or should we go a step further and think:

• What kind of intelligent collaborative ecosystem do we hope to build?

• In this ecosystem, how to balance efficiency and safety, autonomy and controllability?

• How can we ensure that human value is not only not weakened, but amplified through the collaboration of AI?

Embrace connections, but don’t get lost in them.

Use intelligence, but always be wise.

This may be the attitude we should have when facing A2A and more disruptive technologies in the future.

The road is long and arduous, but I will search for it up and down.

Let’s encourage each other!